Harden trading, training, and monitoring

tools/windows_training_agent.py

@@ -64,6 +64,11 @@ def poll_once(args: argparse.Namespace, repo_root: Path, runtime_dir: Path, log_
     try:
         run_retrain(args, job_id, job, repo_root, log_path)
         summary = read_json(runtime_dir / "torch_retrain_guard.json")
+        if summary.get("accepted") is not True:
+            raise RuntimeError(
+                "candidate rejected by untouched-holdout guard: "
+                + str(summary.get("reason") or "validation failed")
+            )
         report_progress(args, job_id, "running", "uploading", 72, "Обучение завершено, загружаю артефакты")
         for name in ARTIFACT_NAMES:
             path = runtime_dir / name
@@ -102,11 +107,14 @@ def run_retrain(args: argparse.Namespace, job_id: str, job: dict[str, Any], repo
         "layers": "-Layers",
         "dropouts": "-Dropouts",
         "epochs": "-Epochs",
+        "holdout_window": "-HoldoutWindow",
     }
     for key, ps_arg in arg_map.items():
         value = parameters.get(key)
         if value not in (None, ""):
             cmd.extend([ps_arg, str(value)])
+    if parameters.get("resume_candidate") is True:
+        cmd.append("-ResumeCandidate")
     log(log_path, "Running retrain: " + " ".join(quote_for_log(part) for part in cmd))
     report_progress(args, job_id, "running", "training", 8, "PyTorch retrain запущен")
     line_count = 0